GT Axis Daily, July 24, 2026

DeepSeek moved its V4 model out of preview status this week, ending months of shifting pricing and behavior that kept cautious companies from trusting it for real work. It's one of the strongest free-to-download ("open-weight," meaning anyone can pull the underlying model and run it themselves) models in the world, and this stabilization is what gets it into actual production systems instead of just demos.

CEO Sundar Pichai told investors the company has kicked off what he called its most ambitious training run yet for Gemini 4, while the current Gemini app just crossed 950 million monthly users. Translation: Google isn't resting on a hit, it's already spending to make the hit obsolete.

Grok Build launched with Grok 4.5 earlier this month, joining a crowded field of AI tools that write and fix code with less hand-holding. For non-developers, the short version: the bar for "AI that can just do the coding task" keeps rising, and it's rising across every major lab at once, not just one.

Alphabet raised its 2026 spending plan to a range of $195 to $205 billion, up from its prior guidance, after cloud revenue jumped 82% for the quarter. That's the company betting the AI buildout pays for itself through cloud customers renting Google's computers and models.

SAP finished buying Prior Labs, a young company that builds AI specialized for spreadsheet-style business data (the kind every company already has, as opposed to the text and images most AI is trained on). SAP is letting it keep operating independently while pouring in funding, essentially buying itself a frontier AI lab focused on the boring, valuable data every enterprise already owns.

Instead of a flat monthly fee for a fixed number of requests, Copilot now runs on "AI Credits," a shared allowance that gets consumed based on how much computing power each request burns. Plan prices didn't change, but heavy users of the priciest AI models will burn through their allowance faster. It's the software world's version of a utility bill replacing a flat-rate plan.

Claude Opus 4.8 is generally available on GitHub Copilot, Amazon's cloud, and Google's cloud all at once, and it's four times less likely than its predecessor to miss bugs in the code it writes. For a business leader, that's the gap between an AI assistant you have to babysit and one you can actually trust with real work.

SAXON Q's new machines run on diamond chips at room temperature, meaning no giant refrigeration units or specialized clean rooms, the stuff that's made quantum computers expensive science projects until now. The first version ships this fall and slides into standard data center hardware, which is the difference between a lab curiosity and something a company could actually buy and use.

MetaX confidentially filed for a Hong Kong stock listing, aiming to raise money by the end of the year to build its own AI chips as Beijing pushes Chinese companies to stop relying on Nvidia's hardware. It's the latest in a wave of Chinese GPU makers rushing to list shares to bankroll years of expensive chip development.

Atoms, built around a warehouse and mining automation business Kalanick bought last year, pulled in the round led by top venture firm Andreessen Horowitz. It's a bet that AI's next big win isn't in chatbots, it's in physically automating dirty, repetitive industrial work, and that the founder who built Uber's logistics machine can do it again in factories and mines.

Brookfield and Bloom Energy expanded their partnership to $25 billion, up from $5 billion, to build on-site fuel cell power plants for data centers that can't wait years for the regular electric grid to catch up. This is the unglamorous part of the AI boom: someone has to physically generate enough electricity to run all these computers, and that's turning into its own multi-billion-dollar industry.